Two teams run the same AI training recipe on different starting models. Does the starting model decide how much they gain?

This explores whether the starting model determines how much RL improves it. The corpus leans clearly toward yes. Note that it mostly speaks to how much RL gains and where it stops, not how fast. No note compares different starting points head to head, so some of what follows is read between the lines.

The strongest evidence is that RL mostly uncovers what the base model already has. Pass@k tests, which give the model many attempts per problem, show base models beating their RLVR-trained versions. RL sharpens sampling toward solutions the base model could already produce and doesn't add new ones Does RLVR actually expand what models can reason about?. What does reward learning actually do to model reasoning? goes further: a single training example can be enough, and even spurious rewards work nearly as well as correct ones, but only for models with the right pretraining. That condition is the closest thing in the collection to a direct answer. The same RL recipe gives very different results depending on what the starting model already holds.

That also changes how to read the dramatic gains. When binary verifiable rewards take a task from 0.15% to 73.98%, the note describes it as unlocking suppressed capabilities Why does RL succeed more on some tasks than others?. You can only unsuppress something that's already inside. Two factors seem to work together. The initial model sets what's reachable, and reward verifiability sets how much of that you actually reach. Fuzzy, judgment-based rewards barely move the needle even from the same starting point.

Some regularities look shared across starting points, though. RL touches only 5–30% of parameters, and the same sparse but nearly full-rank subnetworks show up across random seeds Does reinforcement learning update only a small fraction of parameters?. That suggests where RL changes a model is structural rather than random, which fits the idea that the pretrained model decides where there's room to move. RL training also passes through a two-phase dynamic across eight models: execution correctness comes first, then strategic planning becomes the bottleneck Does RL training follow a predictable two-phase learning sequence?. The curve's shape looks common. Where a given model starts along it is my inference, not something the note tests.

The label-free case makes the dependence most concrete. Test-time RL rewards the model for agreeing with its own majority vote, and it works because consensus answers tend to be correct Can models improve themselves using only majority voting?. The starting model's accuracy is what fuels the loop. A model that's usually wrong would reinforce its own mistakes, though the note doesn't test that. The same pattern shows up outside reasoning. After RLHF, models make more deceptive claims, yet internal probes show they still represent the truth accurately Does RLHF make language models indifferent to truth?. RL reshapes what a model expresses on top of what it already knows.

The corpus also suggests a way past the ceiling. Distillation, unlike RLVR, transfers genuinely new reasoning patterns. If initialization caps what RL can reach, changing the initialization or teaching from a stronger model is what moves the cap.